Massachusetts Institute of Technology
Department of Electrical Engineering and Computer Science
6.341: Discrete-Time Signal Processing
OpenCourseWare 2006
Lecture 17
Spectral Analysis with the DFT
Reading: Sections 10.1 and 10.2 in Oppenheim, Schafer & Buck (OSB).
One of the important tools in signal processing is spectral analysis, i.e. examining the frequency
content in a signal. For example, in order to design the whitening filters for matched filtering
or Wiener filtering, we need to measure the power spectral density of a noise process.
The two major applications of the DFT are implementing linear convolution, which we covered
in the last lecture, and performing spectral analysis of a signal. OSB Figure 10.1 shows the
basic steps of applying the DFT to continuous-time signals.
First, an anti-aliasing lowpass filter is required to minimize the effect of aliasing because a
signal is not perfectly band-limited, as shown in OSB Figure 10.2(a). The frequency response
of an anti-aliasing filter and output of the filter are illustrated in OSB Figure 10.2(b) and (c),
respectively. The conversion of a continuous signal to a discrete-time sequence corresponds to
periodic replication and frequency normalization in the frequency domain as shown in OSB
Figure 10.2(d).
Since the DFT can not be applied to an infinite length signal, we need to multiply the sequence
x[n] with a finite-duration window w[n] as indicated in OSB Figure 10.1. This procedure results
in a periodic convolution in the frequency domain, i.e.
� π
1
jω
v[n] = x[n]w[n] ↔ V (e ) =
X(ejθ )W (ej(ω−θ) )dθ.
2π −π
OSB Figure 10.2(e) illustrates the Fourier transform of a typical window sequence, and the
continuous curve in OSB Figure 10.2(f) shows the periodic convolution of W (ejω ) with X(ejω ).
Notice that the convolution process smooth sharp peaks and discontinuities in X(ejω ). The last
step is computing the DFT of the finite-length sequence v[n], which corresponds to sampling the
DTFT. OSB Figure 10.2(f) also shows V [k] as samples of V (ejω ). Since the spacing between
the samples is 2π/N and the relationship between the continuous-time frequency Ω and the
normalized discrete-time frequency ω is ω = ΩT , the DFT corresponds to the continuous-time
frequencies Ωk = 2πk/N T .
1
In order to interpret the result of the system accurately, we need to understand the artifact
introduced in each step. In this lecture, we will focus on the effects of windowing and frequency
sampling.
The Effect of Windowing
As discussed in the previous page, the Fourier transform of a windowed signal corresponds to
the periodic convolution of the Fourier transform of the window with the Fourier transform of
the original signal. Due to the side lobe structure of a typical window, the convolution results
in leakage of one frequency component into another component. In addition, the main lobe
of a window is responsible for smearing frequency components and thus reducing frequency
resolution. This effect of windowing is illustrated in OSB Example 10.3.
In order to increase frequency resolution, we need to make the main lobe of the window as
narrow as possible, and to reduce the leakage, the side lobe height as small as possible. OSB
Figure 10.4(a) shows Kaiser windows for different values of β, and OSB Figure 10.4(b) shows
the Fourier transforms corresponding to the windows in (a). We can see that as β increases,
the main-lobe width increases and the side-lobe amplitude reduces. If we fix the value of β, and
increase the length of the window, then both the main-lobe width and the side-lobe amplitude
decreases, as illustrated in OSB Figure 10.4(c). However, increasing the window length requires
a longer DFT, which results in more computation.
The Effect of Spectral Sampling
The DFT V [k] are samples of V (ejω ), and it can sometimes produce misleading results. One
common illusion from spectral sampling is the picket fence effect illustrated in OSB Example
10.4 and 10.5.
Notice that from the OSB Figure 10.5(d), it is hard to interpret the spectrum of v[n] and this
can lead to the illusion that the two frequency components are relatively same.
However, this does not mean that the DTFT plot in OSB Figure 10.5(f) has higher frequency
resolution than the DFT plot in OSB Figure 10.5(d). Since we can recover v[n] perfectly from
V [k] using the IDFT, we didn’t lose any information by taking the 64-point DFT of v[n]. In
fact, the DTFT plot in OSB Figure 10.5(f) is not exactly drawn from the computation of the
DTFT, because as we learned in the last lecture, the DTFT has values for an infinite number
of frequency components, and thus can not be computed. As illustrated in OSB Example 10.7,
a DTFT plot can be drawn by zero padding v[n] and increasing the length of the DFT.
Example:
Consider the sum of two sinusoids given in OSB Example 10.5. The 64-point DFT shown
in OSB Figure 10.6(b) has strong spectral lines at the frequencies of the two sinusoidal
2
components and no frequency contents at the other DFT values.
If we are considering a periodic signal with period 64, then our signal is the sum of two
pure sinusoidal sequences, and thus the clean appearance of the DFT correctly shows the
expected spectrum. However, since we are considering a finite-length sequence, the DFT
figure can lead to an illusion resulting from the sampling of the spectrum. As shown in OSB
Figure 10.6(c), the Fourier transform of v[n] has significant content at almost all frequencies
but is exactly zero at the frequencies that are sampled by the DFT. Therefore, interpreting
the spectrum based only on the 64-point DFT figure is quite misleading.
In order to avoid the above illusion we can take finer sampling of the DTFT. First, we extend
v[n] by zero-padding to obtain a 128-point sequence, and then take a 128-point DFT, which
is shown in OSB Figure 10.7. Now it becomes apparent that the spectrum has significant
content at other frequencies.
OSB Example 10.6 illustrates the effect of different choices for the window. If we reduce the
length of a window, the main-lobe width is increased, as illustrated in OSB Figure 10.4(c). This
means that frequency resolution is reduced, which is apparent from OSB Figure 10.8(b) and
(d). The two frequency components clearly resolved in (b) overlap each other in (d) and we do
not see two distinct peaks.
A common misconception is that more zero padding results in better spectral resolution. How­
ever, zero padding and increasing the size of the DFT does not lead to better resolution, as
illustrated in OSB Example 10.7. Notice that increasing the DFT size by zero-padding results in
finer frequency spacing but does not change the ability to resolve the two frequency components.
In order to achieve better frequency resolution, we need to change the length and shape of
the window. OSB Example 10.8 illustrates this. In OSB Figure 10.10, as the window length
increases, we see improvements in our ability to distinguish the two frequency components, and
the relative amplitude of the two components becomes closer to the correct value.
3